5 ms·

Python: where relying on “lol magic” is acceptable thing to have in production code. At least “magic” in languages like Haskell has the dignity to be backed up

Python: where relying on “lol magic” is acceptable thing to have in production code. At least “magic” in languages like Haskell has the dignity to be backed up by a type system with more consistency than a wet noodle. Python can’t seem to decide whether it wants to be a production grade language or easy-for-beginners and keeps implementing pointless features (looking at you walrus operator and weird version or pattern matching) instead of fixing its actual problems like the tire-fire that is packaging and performance. “But Python has never been about performance” someone will say, and that’s fine, but it’s increasingly becoming a language that isn’t any more ergonomic to use than its peers, with worse packaging and deployment, useful tools (type hunting) that the community seemingly refuses to adopt widespread and its slower on top of all that? Why would you actually bother?

PartiallyTyped 6y ago
> But Python has never been about performance I have been burnt by this. As much as I like python's "expressiveness" so to speak, it is too slow for my needs; the ML community gravitated towards that because it is simple to implement and test large models because C++/C/Cuda are doing the lifting. However, in my flavour/field of ML, python is often the bottleneck because the sampling/simulations are glued with it and the need to sample 10s of millions of steps ends up with possibly billions of function calls in python.
- dunefox 6y ago
  You might take a look at Julia with PyCall and RCall, if you haven't already. I find it solves a lot of headaches I have with Python while allowing me to leverage the ecosystem.
  - PartiallyTyped 6y ago
    Yes, I am currently learning Julia. My issue with it is that the documentation is rather sparse or lacking, but I am getting the hang of it. On the positive side, it appears that the jitter works quite well and fast arrays as first class citizens make me happy.
    - dunefox 6y ago
      There's this, could be helpful for basic Julia: https://benlauwens.github.io/ThinkJulia.jl/latest/book.html https://benlauwens.github.io/ThinkJulia.jl/latest/book.html

PartiallyTyped 6y ago

> But Python has never been about performance I have been burnt by this. As much as I like python's "expressiveness" so to speak, it is too slow for my needs; the ML community gravitated towards that because it is simple to implement and test large models because C++/C/Cuda are doing the lifting. However, in my flavour/field of ML, python is often the bottleneck because the sampling/simulations are glued with it and the need to sample 10s of millions of steps ends up with possibly billions of function calls in python.

dunefox 6y ago
You might take a look at Julia with PyCall and RCall, if you haven't already. I find it solves a lot of headaches I have with Python while allowing me to leverage the ecosystem.
- PartiallyTyped 6y ago
  Yes, I am currently learning Julia. My issue with it is that the documentation is rather sparse or lacking, but I am getting the hang of it. On the positive side, it appears that the jitter works quite well and fast arrays as first class citizens make me happy.
  - dunefox 6y ago
    There's this, could be helpful for basic Julia: https://benlauwens.github.io/ThinkJulia.jl/latest/book.html https://benlauwens.github.io/ThinkJulia.jl/latest/book.html